In plain words
Video generators can make long videos a section at a time, using each section to make the next. Small errors in earlier sections can build up and make later parts look worse.
Clean Forcing adds a small learned correction to a video model while keeping its main model fixed. On 128 prompts, it reduced an image-quality decline score in 50-second videos, even when trained without real video clips. The score tracks one aspect of quality, not everything viewers may notice.
The research
When a video model generates one chunk after another, each new chunk depends on its imperfect earlier output. Those errors can accumulate and degrade longer videos. Can a correction learned for the model’s generated histories reduce that drift while leaving the base model’s weights frozen?
Clean Forcing learns a small correction for a frozen video model, targeting prediction errors caused by the imperfect chunks it generates itself.
Researchers ask whether the mismatch between clean training histories and a model’s imperfect generated histories can be corrected without retraining the full video model. Clean Forcing learns a small correction to the frozen model’s predictions under generated histories.
On 128 held-out prompts, the corrector lowered the mean image-quality drift score over 50-second videos, both with no real training clips and with 40 clips. The matched prompts and seeds connect the change to adding the corrector under this evaluation protocol.
The main result measures an automated image-quality score on 128 prompts under a 50-second generation protocol; it does not capture every aspect of video quality.
Stronger drift reduction generally came with more scene anchoring, and the method still needs a new corrector for each checkpoint or domain.
Analysis
Deep
- Original paper authors
- Wenqing Wang, Joonghyuk Shin, Jonathan Tremblay, Chan Hee Song, Yun Fu
- Submitted to arXiv
- License
- CC BY 4.0